Long-read single-cell sequencing reveals expressions of hypermutation clusters of isoforms in human liver cancer cells

  1. Department of Pathology
  2. High Throughput Genome Center
  3. Pittsburgh Liver Research Center
  4. University of Pittsburgh, 3550 Terrace Street, Pittsburgh, PA 15261; Element Biosciences, Inc, 10055 Barnes Canyon Road, Suite 100, San Diego, CA 92121
  5. Biostatistics


  • Reviewing Editor
    Thomas Gingeras
    Cold Spring Harbor Laboratory, Cold Spring Harbor, United States of America
  • Senior Editor
    David James
    University of Sydney, Sydney, Australia

Reviewer #1 (Public Review):

In the manuscript "Long‐read single‐cell sequencing reveals expressions of hypermutation clusters of isoforms in human liver cancer cells", S. Liu et al present a protocol combining 10x Genomics single-cell assay with Element LoopSeq synthetic long-read sequencing to study single nucleotide variants (SNVs) and gene fusions in Hepatocellular carcinoma (HCC) at single‐cell level. The authors were the first to combine LoopSeq synthetic long‐read sequencing technology and 10x Genomics barcoding for single cell sequencing. For each cell and each somatic mutation, they obtain fractions of mutated transcripts per gene and per each transcript isoform. The manuscript states that these values (as well as gene fusion information) provide better features for tumor-normal classification than gene expression levels. The authors identified many SNVs in genes of the human major histocompatibility complex (HLA) with up to 25 SNVs in the same molecule of HLA‐DQB1 transcript. The analysis shows that most mutations occur in HLA genes and suggests evolution pathways that led to these hypermutation clusters. Yet, very little is said about novel isoforms and alternative splicing in HCC cells, differences in isoform ratio between cells carrying different mutations, or diversity of alternative isoforms across cells. While the manuscript by Liu et al. presents a promising combination of technologies, it lacks significant insights, a comprehensive introduction, and has significant problems with data description and presentation.

Major comments:

1. The introduction section is scarce. It lacks description of important previous works focused on clustered mutations in cancers (for example, PMID35140399), on deriving the process of cancer development through somatic evolution (PMID32025013, from single cell data PMID32807900). Moreover, some key concepts e.g. mutational gene expression and mutational isoform expression are not defined. The introduction and the abstract contain slang expressions e.g. "protein mutation', a combination of terms I teach my students not to use.

2. In the results section, to select the mutations of interest, the authors apply UMAP dimensionality reduction to the mutation isoforms expression and cluster samples in UMAP space, then select the mutations that are present only in one cluster, then apply UMAP to the selected mutations only and cluster the samples again. The motivation for such a procedure seems unclear, could it be replaced with a more straightforward feature selection?

3. As I understand, the first "mutated isoform"-based UMAP clustering was built from expression levels of 205 "mutational isoforms". What was the purpose and outcome of the second "mutated isoform"-based UMAP clustering (Figure 2E)? In the manuscript the authors just describe the clusters and do not draw any conclusions or use the results of the clustering anywhere further.

4. The authors just cluster the data three times based on expression levels of different sets of "mutational isoforms" and describe the clusters. What do we need to gather from these clustering attempts besides the set of 113 mutations used for further analysis? What was the point of the re-clusterings? Did the authors observe improvement of the classification at each step?

5. The alignment of short reads generated from hypermutated transcriptomes is non-trivial. The proposed approach could address the issue without need for whole genome sequencing and offer insights about the cancer development through somatic evolution. Why didn't the authors use modern phylogenetic approaches in the "Evolution of mutations in HLA molecules" section or at least utilize the already performed clustering to infer cell lineages?

6. I am not sure I understood the definition of "mutated gene expression levels" and "mutated isoform expression levels" in the "Mutational gene expression and fusion transcript enhanced transcriptome clustering of benign hepatocytes and HCC" section. The authors mention that gene lists included all the isoforms within the same range of standard deviation. If I understand it correctly, they are equal if there is only one expressed transcript isoform. In that case, this overlap is not surprising at all.

7. "To investigate the roles of gene expression alterations that were not accompanied with isoform expression changes, UMAP analyses were performed based on the non‐overlapped genes." Venn diagrams (Sup Figure 8) show that there are much less "non-overlapped genes" than "genes that showed both gene and isoform level changes" for each SD threshold (for example, for SD>=0.8 59 vs 275). Could that be the reason why clustering based on the former group is worse i.e the cancer and normal cells are separated less clearly?

Reviewer #2 (Public Review):

In the present study, Liu et al present an analysis of benign and HCC liver samples which were subjected to a new technology (LOOP-Seq) and paired WES. By integrating these data, the authors find isoforms, fusions and mutations which uniquely cluster within HCC samples, such as in the HLA locus, which serve as candidate leads for further investigation. The main appeal of the study is in the potential of LOOP-Seq as a method to present isoform-resolved data without actually performing long-read sequencing. While this presents an exciting new method, the current study lacks systematic comparisons with other technologies/data to test the robustness, reproducibility and utility of LOOP-Seq. Further, this study could be further improved by giving more physiologic context and examples from the analyses, thus providing a new resource to the HCC community. A few suggestions based on these are below:

A primary consideration is that this seems to be the first implementation of LOOP-Seq, where the technology, while intriguing, has not been evaluated systematically. It seems like a standard 10x workflow is performed, where exons are selectively pulled down and amplified. Subsequent ultra-deep sequencing is assumed to give isoform-resolution of the sc-seq data. To demonstrate the utility of the approach it would benefit the study to compare the isoform-resolved results with studies where long-read sequencing was actually performed (ex: https://journals.lww.com/hep/Fulltext/2019/09000/Long_Read_RNA_Sequencing_Identifies_Alternative.19.aspx, https://www.jhep-reports.eu/article/S2589-5559(22)00021-0/fulltext, https://journals.plos.org/plosgenetics/article?id=10.1371/journal.pgen.1010342). Presumably, a fair amount of overlap should occur to justify the usage.

Related to this point, the sc-seq cell types and benign vs HCC genes should be compared with the wealth of data available for HCC sc-seq (https://www.nature.com/articles/s41467-022-32283-3, https://www.nature.com/articles/s41598-021-84693-w). These seem to be important to benchmark the technology in order to demonstrate that the probe-based selection and subsequent amplification does not bias cell type definition and clustering. In particular, https://www.nature.com/articles/s41586-021-03974-6 seems quite relevant to compare mutational landscapes from the data.

From the initial UMAP clustering, it will be important to know what the identities are of the cells themselves. Presumably there is quite a bit of immune cells and hepatocytes, but without giving identities, downstream mechanistic interpretation is difficult.

In general, there are a fair amount of broad analyses, such as comparisons of hierarchical clustering of cell types, but very little physiologic interpretations of what these results mean. For example, among the cell clusters from Fig 6, knowing the pathways and cell annotations would help to contextualize these results. Without more biologically-meaningful aspects to highlight, most of the current appeal for the manuscript is dependent on the robustness of LOOP-seq and its implementation.

Many of the specific analyses are difficult and the methods are brief. Especially given that this technology is new and the dataset potentially useful, I would strongly recommend the authors set up a git repository, galaxy notebook or similar to maximize utility and reproducibility

The authors claim that clustering between benign and HCC samples was improved by including isoform & gene (Suppl fig 8). This seems like an important conclusion if true, especially to justify the use of long-read implementation. Given that the combination of isoform + gene presents ~double the number of variables on which to cluster, it would be important to show that the improved separation on UMAP distance is actually due to the isoforms themselves and not just sampling more variables from either gene or isoform

SQANTI implementation to identify fusions relevant for the HCC/benign comparison. How do the fusions compare with those already identified for HCC? These analyses can be quite messy when performed on WES alone so it seems that having such deep RNA-seq would improve the capacity to see which fused genes are strongly expressed/suppressed. This doesn't seem as evident from current analysis. There are quite a bit of WES datasets which could be compared: https://www.nature.com/articles/ng.3252, https://www.nature.com/articles/s41467-018-03276-y

Figure 4 is fairly unclear. The matrix graphs showing gene position mutations are tough to interpret and make out. Usually, gene track views with bars or lollipop graphs can make these results more readily interpretable. Also, how Figure 4 B infers causal directions from mutations is unclear.

Reviewer #3 (Public Review):

The Liu, et al. manuscript focuses on the interesting topic of evaluating in an almost genome-wide-scale, the number of transcriptional isoforms and fusion gene are present in single cells across the annotated protein coding genome. They also seek to determine the occurrences of single nucleotide variations/mutations (SNV) in the same isoform molecule emanating from the same gene expressed in normal and normal and hepatocellular carcinoma (HCC) cells. This study has been accomplished using modified LoopSeq long‐read technology (developed by several of the authors) and single cell isolation (10X) technologies. While this effort addresses a timely and important biological question, the reader encounters several issues in their report that are problematic.:

  1. Much of the analysis of the evolution of mutations results and the biological effects of the fusion genes is conjecture and is not supported by empirical data. While their conclusions leave the reader with a sense that the results obtained from the LoopSeq has substantive biological implications. However, they are extended interpretations of the data. For example: The fusion protein likely functions as a decoy interference protein that negatively impacts the microtubule organization activity of EML4.(pg 9)... and other statements presented in a similar fashion.

  2. LoopSeq has the advantage of using short read sequencing analyses to characterize the exome capture results and thus benefits from low error rate compared to standard long-read sequencing techniques. However, there is no evidence obtained from standard long read sequencing that the isoforms observed with LoopSeq are obtained with parallel technologies such as long read technologies. It is not made clear how much discordance there is in comparing the LoopSeq results are with either PacBio or ONT long read technologies.

  3. There is no proteome evidence (empirically derived or present in proteome databases) from the HCC and normal samples that confirms the presence or importance of the identified novel isoforms, nor is there support that indicate that changes in levels HLA genes translate to effects observed at the protein level. Since the stability and transport differences of isoforms from the same gene are often regulated at the post-transcriptional level, the biological importance of the isoform variations is unclear.

  4. It is unclear why certain thresholds were chosen for standard deviation (SD) <0.4 (page 5), SD >1.0 (pg 11).

  5. HLA is known to accumulate considerable somatic variation. Of the many non-immunological genes determined to have multiple isoforms what are the isoform specific mutation rates in the same isoform molecule? Are the HLA genes unique in the number of mutations occurring in the same isoform?

  1. Howard Hughes Medical Institute
  2. Wellcome Trust
  3. Max-Planck-Gesellschaft
  4. Knut and Alice Wallenberg Foundation